Classification, inference and segmentation of anomalous diffusion with recurrent neural networks
Aykut Argun, Giovanni Volpe, Stefano Bo
J. Phys. A: Math. Theor. 54 294003 (2021)
Countless systems in biology, physics, and finance undergo diffusive dynamics. Many of these systems, including biomolecules inside cells, active matter systems and foraging animals, exhibit anomalous dynamics where the growth of the mean squared displacement with time follows a power law with an exponent that deviates from 1. When studying time series recording the evolution of these systems, it is crucial to precisely measure the anomalous exponent and confidently identify the mechanisms responsible for anomalous diffusion. These tasks can be overwhelmingly difficult when only few short trajectories are available, a situation that is common in the study of non-equilibrium and living systems. Here, we present a data-driven method to analyze single anomalous diffusion trajectories employing recurrent neural networks, which we name RANDI. We show that our method can successfully infer the anomalous exponent, identify the type of anomalous diffusion process, and segment the trajectories of systems switching between different behaviors. We benchmark our performance against the state-of-the art techniques for the study of single short trajectories that participated in the Anomalous Diffusion (AnDi) challenge. Our method proved to be the most versatile method, being the only one to consistently rank in the top 3 for all tasks proposed in the AnDi challenge.
Deep learning for microscopy, optical trapping, and active matter
(online at) TU-Darmstadt, Germany
18 June 2021, 14:00 CEST
After a brief overview of artificial intelligence, machine learning and deep learning, I will present a series of recent works in which we have employed deep learning for applications in photonics and active matter.
In particular, I will explain how we employed deep learning to enhance digital video microscopy, to estimate the properties of anomalous diffusion, to characterize microscopic force fields, to improve the calculation of optical forces, and to characterize nanoparticles.
Finally, I will provide an outlook for the application of deep learning in photonics and active matter.
The disputation took place at 9 a.m. digitally via Zoom. A link to the Zoom meeting was published the day before dissertation on the GU website.
Title: Deep Learning Applications – From image analysis to medical diagnosis
Deep learning is a subcategory of machine learning and artificial intelligence. Instead of using explicit rules to perform a desired task as in standard algorithmic approaches, machine-learning algorithms autonomously learn from data to determine the rules for the task at hand. The idea of deep learning has been around since the 1950s but was for a long time limited by available computational power and amount of training data. Once overcome these problems, in recent years, deep learning has made great advances in solving various problems.
In this thesis, I show how deep learning can be applied in image analysis and medical diagnosis, while outperforming standard algorithmic methods and simpler machine-learning methods. I begin with showing that a convolutional neural network trained with simulated particle images is able to track experimental single particles, even in poor illumination conditions. I then show how this inspired the development of an all-in-one software to design, train and validate deep-learning solutions for digital microscopy, from particle tracking and characterization in 2D and 3D to the segmentation, characterization and counting of biological cells and image transformation. I show that this software package can be further used to develop a generative adversarial neural network to virtually stain brightfield images of cells, replacing the traditional chemical staining for a downstream analysis of biological features. I then move on from applications in microscopy and image analysis to show the potential of deep learning in medical diagnosis. I show that dense neural networks perform better than simpler machine-learning algorithm and the clinical standard in the diagnosis of a genetic disease and in the prediction of short- and long-term morbidity in patients with congenital-heart-disease. At last, I have shown that a neural network- powered strategy for testing and isolating individuals adapts to the parameters of a disease outbreak achieves an epidemic containment.
The interdisciplinary nature of the work in this thesis has allowed the application of new technologies developed in the field of physics to solve problems in the fields of biology and biomedicine, as well as overcoming barriers for the continued revolutionization of deep learning in microscopy.
Olle Fager defended his Master thesis in MPCAS at the Chalmers University of Technology on 15 June 2021. Congrats!
Title: Real-Time Multi-Object Tracking and Segmentation with Generated Data using 3D-modelling
Multi-Object Tracking and Segmentation (MOTS) is an important branch of computer vision that has applications in many different areas. In recent developments these methods have been able to reach favorable speed-accuracy trade-offs, making them interesting for real-time applications. In this work different deep learning based MOTS methods have been investigated with the purpose of extending the DeepTrack framework with real-time MOTS capabilities. Deep learning methods rely heavily on the data on which they are trained. The collection and annotation of the data can however be very time-consuming. Therefor, a pipeline is developed and investigated that automatically produces synthetic data by utilizing 3D-modelling. The most accurate tracker achieves a MOTSA score of 94 and the tracker with the best speed-accuracy trade-off achieves a MOTSA score of 88. It is also observed that satisfactory results can be achieved in most situations with a quite general data generation pipeline, indicating that the developed pipeline could be used in different scenarios.
Name of the master programme: MPCAS – Complex Adaptive Systems Supervisor: Giovanni Volpe Examiner: Giovanni Volpe, Department of Physics, University of Gothenburg Opponent: Arianit Zeqiri and Morad Mahmoudyan
Santhosh Shivan Gurumurthy joined the Soft Matter Lab on 15 June 2021.
Santhosh is a master student in Complex Adaptive Systems at Chalmers University of Technology.
During his time at the Soft Matter Lab, he will focus on using Deeptrack tools to simulate microscopic particles, track them using various object detection algorithms, and predict their trajectory with time using graph neural networks (GNN).
Aykut Argun defended his Ph.D. thesis on June 14, 2021, at 2 pm CEST. Congrats!
The details of the presentation can be found below. The link to the webinar is announced on the faculty website.
Title: Thermodynamics of microscopic environments: From anomalous diffusion to heat engines.
Unlike their macroscopic counterparts, microscopic systems do not evolve deterministically due to the thermal noise becoming prominent. Such systems are subject to fluctuations that can only be studied within the framework of stochastic thermodynamics. Within the last few decades, the development of stochastic thermodynamics has lead to microscopic heat engines, nonequilibrium relations and the study of anomalous diffusion and active Brownian motion. In this thesis, I experimentally show that the non-Boltzmann statistics emerge in systems that are coupled to an active bath. These non-Boltzmann statistics that result from correlated active noise also disturb the nonequilibrium relations. Nevertheless, I show that these relations can be recovered using an effective potential approach. Next, I demonstrate an experimental realization of a microscopic heat engine. This engine is referred to as the Brownian gyrator, which is coupled to two different heat baths along perpendicular directions. I show that when confined into an elliptical trap that is not aligned with the temperature anisotropy, the Brownian particle is subject to a torque due to the symmetry breaking. This torque creates an autonomous engine whose direction and amplitude can be controlled by tuning the alignment of the elliptical trap. Then, I show that the force fields acting on Brownian particles can be calibrated using a data-driven method that outperforms the existing calibration methods. More importantly, I show that this method, named DeepCalib, can calibrate non-conservative and time-varying force fields that no standard calibration methods exist. Finally, I show that a similar machine-learning-based approach can be used to characterize anomalous diffusion from single trajectories. This method, named RANDI, is very versatile and performs very well in various tasks including classification, inference and segmentation of anomalous diffusion. The work presented in this thesis presents novel experiments that advance microscopic thermodynamics as well as newly developed methods that open up new possibilities in analyzing stochastic trajectories. These findings increased the scientific knowledge at the nexus between microscopic thermodynamics, anomalous diffusion, active matter and machine learning.
Supervisor: Giovanni Volpe Co-supervisors: Joakim Stenhammar, Mattias Goksör Examiner: Bernhard Mehlig Opponent: Juan M. R. Parrondo Committee: Monika Ritsch-Marte, Sabine H. L. Klapp, Édgar Roldán
Screenshots from Aykut Argun’s PhD Thesis defense.
Dendritic spines are lost in clusters in patients with Alzheimer’s disease
Mite Mijalkov, Giovanni Volpe, Isabel Fernaud-Espinosa, Javier DeFelipe, Joana B. Pereira, Paula Merino-Serrais
Sci. Rep. 11, 12350 (2021)
Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by a deterioration of neuronal connectivity. The pathological accumulation of tau protein in neurons is one of the hallmarks of AD and has been connected to the loss of dendritic spines of pyramidal cells, which are the major targets of cortical excitatory synapses and key elements in memory storage. However, the detailed mechanisms underlying the loss of dendritic spines in patients with AD are still unclear. Here, comparing dendrites with and without tau pathology of AD patients, we show that the presence of tau pathology determines the loss of dendritic spines in blocks, ruling out alternative models where spine loss occurs randomly. Since memory storage has been associated with synaptic clusters, the present results provide a new insight into the mechanisms by which tau drives synaptic damage in AD, paving the way to memory deficits by altering spine organization.
Job assignments: Possibility 1: Work at the development of DeepTrack 2.0 (https://aip.scitation.org/doi/10.1063/5.0034891 and https://github.com/softmatterlab/DeepTrack-2.0) in close cooperation with the existing developers’ team at Soft Matter Lab. DeepTrack 2.0 is a software framework we are developing to perform quantitative digital video microscopy using deep learning. The PhD student will actively contribute to the technical design and implementation of the components of DeepTrack. Precisely, the task involves the development, testing and application of optical simulation pipelines for training deep learning networks, as well as the development of collaborative projects with other groups interested in using DeepTrack. Possibility 2: Build and operate small robots to study swarm robotics with embedded intelligence. The task involves all stages of design, fabrication, testing and programming the robots for different experiments. The robot models (inspired by the Kilobots https://ssr.seas.harvard.edu/kilobots) are small programmable units capable of motion, short-range communications, neighbor detection and more.
Appointment procedure Please apply online
The application shall include:
– Cover letter with an explanation of why you apply for the position
– CV including scientific publications
– Copy of exam certificate
– Two referees (name, telephone number, relation)
For the required Qualifications, Eligibility, Assessment criteria, Employment, see the link to the announcement below.
Agaton Fransson defended his Master thesis in MPCAS at the Chalmers University of Technology on 4 June 2021. Congrats!
Title: Tracking plankton using neural networks trained on simulated images
Softwares to track particles often use algorithmic approaches to detect particles and to create tracks using the found positions, requiring human fine-tuning of parameters to achieve sought-for results. This can be time consuming and difficult, while also creating opportunities for human error and bias. With the developments of computational power and machine learning techniques such as deep learning, data driven approaches have made their way into many fields of science. Barriers preventing advances of such methods are the lack of available training data within a field and the level of proficiency required to create custom machine learning solutions. DeepTrack 2.0 is a software providing us with means to simulate digital microscopy images, build and train neural networks such as U-nets. In this paper DeepTrack 2.0 is utilized and built on to fit the needs of marine biologists when tracking plankton. Here I show that DeepTrack 2.0 provides us with the tools necessary to detect and track different types of plankton filmed in a variety of conditions with performance on par with and with the potential to outperform conventional tracking softwares. I also show that for plankton in a messy environment moving uniformly a network trained to detect motion rather than a shape proves more successful. These results demonstrate the versatility of deep learning methods and the potential of training networks on simulations for applications on real data, as is the case for marine biologists studying plankton. They also show the impact the structure of the training data has on the nature of the network.
Name of the master programme: MPCAS – Complex Adaptive Systems Supervisor: Giovanni Volpe, Daniel Midtvedt Examiner: Giovanni Volpe, Department of Physics, University of Gothenburg Opponent: Kevin Rylander
Improving epidemic testing and containment strategies using machine learning. Laura Natali, Saga Helgadottir, Onofrio M. Maragò, Giovanni Volpe. Submitted to SDS2020 Date: 11 June Time: 16:15 (CEST)
Containment of epidemic outbreaks entails great societal and economic costs. Cost-effective containment strategies rely on efficiently identifying infected individuals, making the best possible use of the available testing resources. Therefore, quickly identifying the optimal testing strategy is of critical importance. Here, we demonstrate that machine learning can be used to identify which individuals are most beneficial to test, automatically and dynamically adapting the testing strategy to the characteristics of the disease outbreak. Specifically, we simulate an outbreak using the archetypal susceptible-infectious-recovered (SIR) model and we use data about the first confirmed cases to train a neural network that learns to make predictions about the rest of the population. Using these prediction, we manage to contain the outbreak more effectively and more quickly than with standard approaches. Furthermore, we demonstrate how this method can be used also when there is a possibility of reinfection (SIRS model) to efficiently eradicate an endemic disease.